Previously on Chiron: Juniper left the previous issue with a growing suspicion that the organization’s AI wasn’t merely assisting documentation—it was subtly reshaping the meaning of clinical work. After realizing that the correction queue had changed without anyone noticing, she returned to the clinic determined to understand what, exactly, the system had begun learning.
Note: All names used in Chiron are fictitious. Additionally, this is the third of eight episodes in which we build a story arc using the same characters. At the end, you will find a character cheat sheet to help keep everyone straight from episode to episode.
Juniper did not sleep well after realizing the correction queue changed the meaning the organization’s improving quality metrics.
That was the first thing she checked when she returned to the clinic Friday morning. Same login screen. Same compliance tiles. Same treatment plan tracker. Same percentage of documentation submitted on time. Same cheerful green indicator beside the quality assurance queue.
Correction requests remained down.
That number had seemed beautiful two weeks earlier. It seemed like evidence the platform was working, clinicians were improving, supervision was becoming more efficient, and that the organization had finally reduced one of the small frictions everyone hated but no one had time to solve.
Now it looked less like a measure and more like a mask.
Juniper sat alone in her office before the clinic fully woke. The hallway lights were still dim. Someone at the front desk had started coffee. The first RBTs would arrive in fifteen minutes, then the building would fill with backpacks, therapy materials, parent questions, session transitions, and the institutional hum of people trying to make complicated work look ordinary.
She opened the first record from six months earlier.
Red comments filled the margins.
Clarify the setting events.
Data do not yet support schedule thinning.
Add caregiver implementation context.
Revise recommendation to distinguish acquisition from generalization.
She opened a current record from the same region, same supervisor group, same clinical program.
No comments.
She opened another.
No comments.
Another.
One minor wording suggestion.
Another.
Approved.
Juniper began building a spreadsheet because when her instincts became too loud, she preferred to put them somewhere observable. She sampled records by clinician, region, program type, supervisor, and date. Then she added columns.
AI recommendation accepted.
AI recommendation modified.
AI recommendation rejected.
Human correction entered.
Clarification requested.
Rationale provided.
She paused over the last column.
It had not been there in the old workflow: Rationale Provided.
She opened another current record and scrolled to the bottom.
Explain why you accepted, modified, or rejected this recommendation.
The clinician had written two sentences.
“The recommendation was accepted due to stable responding across three consecutive sessions and no observed procedural drift during implementation. Caregiver training frequency will remain unchanged until generalization is observed across evening routines.”
Juniper copied the response into her spreadsheet.
Then she opened another record.
Explain why you accepted, modified, or rejected this recommendation.
Another response.
“The recommendation was modified because clinic data did not align with caregiver report and the home routine includes variables not captured in the session summary. Additional caregiver probes are recommended before reducing support.”
Juniper copied that too.
By eight-thirty, the clinic had begun moving around her. Voices passed her door in fragments.
"Can you grab the token board?"
"Parent canceled at ten."
"Who has room three?"
"Can you laminate these for me today?"
Juniper kept reading.
The corrections had not disappeared.
They had changed form.
The old workflow displayed clinical disagreement as margin comments, revision requests, and returned documents. The new workflow absorbed disagreement into rationale fields, confidence ratings, and structured prompts that seemed to close the loop before a supervisor ever saw the friction.
At nine-fifteen, Rowan appeared in her doorway holding a paper cup of coffee and the look of someone who had already decided not to pretend the morning was normal.
"You found something," he said.
Juniper did not look up.
"I found a different question…It’s not good."
She turned the monitor toward him.
Rowan stepped inside and closed the door behind him.
He read silently. Then he leaned closer.
"Rationale fields?"
"Across treatment modifications, caregiver coaching plans, supervision summaries, and AI recommendations."
"When did that start?"
"Four weeks ago for pilot regions. Two weeks ago across the organization."
"Was there a training?"
"A workflow update."
Rowan kept reading.
"These are good rationales."
"Well, some of them are."
He pointed at the screen.
"What happens to them?"
Juniper clicked into the platform documentation. The help page opened with its usual soft language.
Rationale responses support transparency, clinical continuity, quality review, and future workflow improvement.
Rowan read the sentence twice. Once to himself. Then once aloud.
"Future workflow improvement."
"That's what it says."
"That's a bit vague, no?"
Juniper leaned back.
"I know."
He looked at her spreadsheet again.
"Does Vale know?"
"Vale knows the workflow changed."
"That's not what I asked."
Juniper closed the help page.
"No. I do not think Vale knows what I am asking yet."
Artificial intelligence (AI) systems do not improve simply by more data being poured into a training dataset. They improve when human-labeled examples help connect conditions, decisions, and consequences. In many AI systems, human responses become especially valuable when they distinguish a useful output from a less useful one. A clinician accepting, editing, rejecting, or explaining an output may create a structured example of professional judgment. That example can be used to evaluate current system performance, improve future workflow, or potentially shape future model behavior, depending on how the organization and vendor have arranged system training.
This matters in behavior analysis because our work often produces comparatively rich records of decision-making. Session notes, prompt changes, functional assessment summaries, treatment modifications, supervision feedback, caregiver coaching plans, and fidelity data do more than document that services occurred. When written well, they capture relations between environmental variables, clinical judgment, and socially meaningful outcomes. That makes behavior analytic documentation useful for continuity of care. It also makes it valuable to any system designed to identify patterns in professional decisions.
For BCBAs and clinical leaders, the practical takeaway is simple: do not evaluate AI tools only by looking at what they generate. Evaluate what they collect. If a system asks clinicians to provide rationales, confidence ratings, corrections, or preferences, leaders should ask where those data go, how they are stored, whether they are identifiable, who can access them, and whether they are used to tune, train, benchmark, or improve future systems. Good documentation can support ethical practice. It can also become infrastructure for collecting data on your professional judgment.
By late morning, Juniper had pulled enough examples to know the pattern was real, but not enough to know what it meant. She did not want to become the kind of clinical leader who mistook discomfort for evidence. The platform might be doing exactly what it claimed. The rationale fields might improve review quality. They might make supervision more efficient. They might reduce vague documentation, clarify clinical decisions, and support continuity when staff changed. In some ways, Juniper could already see the benefits.
A weak rationale was easier to identify than a vague note.
A modified recommendation became easier to audit when the clinician explained the variable the model missed.
A rejected output created a record of disagreement.
Those were real improvements.
That was the irritating part.
Near lunch, she found Mira in the small office shared by the newer BCBAs. Mira had three tabs open, a half-finished salad beside her keyboard, and the posture of someone who had planned to eat but accidentally continued working.
Juniper knocked lightly on the doorframe.
"Do you have five minutes?"
Mira glanced at the screen.
"Is this five minutes or five clinical-leader minutes?"
"Probably the latter."
Mira closed the salad container and turned in her chair.
Juniper stepped inside.
"I want to ask about the new rationale prompts."
"The AI ones?"
Mira nodded.
"They're fine. A little repetitive, but fine."
"Do they change how you write?"
Mira considered the question longer than Juniper expected.
"Yes. But mostly in a good way."
"How?"
"They force me to be more precise. If I accept a recommendation, I have to say why. If I change it, I have to identify the variable. It is kind of like having a supervisor ask the obvious question before I submit."
"Does it feel like supervision?"
"Sometimes. Not personal supervision. More like a checklist with better language."
Juniper nodded.
"Do you ever write the rationale you think the system wants?"
Mira smiled faintly.
"That is a slippery slope."
"I know."
Mira looked at her screen.
"Sometimes. Not because I am trying to game it. But you can tell when certain responses go through more cleanly. Mention the data pattern. Mention implementation fidelity. Mention context. Do not sound too certain. Do not sound too vague. Then it accepts it."
"And if you explain it the way you would to another clinician?"
"Depends on the clinician."
Juniper waited.
Mira sighed.
"Sometimes it asks for clarification."
"What kind?"
Mira clicked into a recent record and scrolled to a returned rationale.
Please specify the contextual variable influencing the clinical decision.
Her original sentence sat above it.
“Family routines have been inconsistent this week, so I am not comfortable thinning caregiver support yet.”
Mira had revised it.
“Caregiver implementation varied across evening routines this week, limiting confidence that current responding will maintain under reduced support. Recommendation modified to continue caregiver coaching frequency pending additional home-context probes.”
Juniper read both versions.
The second was better documentation.
The first sounded more like a clinician.
"Which one do you prefer?" Juniper asked.
Mira looked at them.
"For the record? The second."
"And for supervision?"
Mira did not answer immediately.
"The first would have started a better conversation."
At one o'clock, Juniper brought the issue to Vale.
Vale's office had always been too clean for the kind of work the organization did. He kept a glass board on one wall with quarterly initiatives written in neat columns. Growth. Retention. Compliance. Access. Quality. None of the words were wrong. They simply seemed too stable for the realities they were meant to organize.
Vale listened while Juniper described the correction queue, the rationale fields, and the early pattern she had seen in the records.
He did not interrupt.
That was one of Vale's strengths. He could listen operationally without making a person feel rushed. It was also one of his weaknesses. Sometimes Juniper could see him turning concerns into implementation categories before the concern had fully become itself.
When she finished, he tapped his pen against the notebook once.
"So the system is asking clinicians to explain their decisions."
"Yes."
"And your concern is that the explanations are replacing corrections?"
"Partly."
"What is the other part?"
"I am not sure yet."
Vale leaned back.
"That makes this difficult."
"I know."
"Because on the surface, this is what we want. More transparency. More structured clinical reasoning. Fewer vague approvals. Better audit trail."
"I agree."
"That sounded painful."
"It was."
He smiled, but only briefly.
"Have you talked to Mercer?"
Juniper looked at him.
"Why would I talk to Mercer?"
"Because this touches data strategy. And because he is going to ask the question everyone else is avoiding."
"Which is?"
Vale closed his notebook.
"If clinicians are producing valuable reasoning every day, why would the organization choose not to learn from it?"
